A new framework called Local2Mesh builds a 3D model of the heart's left ventricle from just a handful of flat MRI slices, no 3D scan required.
Researchers describe Local2Mesh in a paper posted to arXiv. The system starts with a template mesh and deforms it to match sparse 2D contours traced from cardiac MRI slices. It first corrects for the fact that those slices are rarely perfectly aligned, a problem the paper calls inter-slice misalignment. A plane-aware Local Router then matches local image features to the nearest points on the template mesh instead of blending everything into one global signal. Tested on two public datasets, M&Ms-2 and ACDC, the method outperformed existing approaches at both reconstructing heart geometry and estimating function, and held up when transferred to a dataset it was not trained on.
That zero-shot transfer is the detail worth noting. Cardiac MRI is usually captured as a stack of flat slices rather than a true 3D volume, because full 3D scans are slower and more expensive, so most reconstruction methods are stuck working with sparse, misaligned data. A technique that generalizes across datasets without needing 3D mesh annotations could cut the manual-labeling burden that has kept this field small.
Still, this is a benchmark result on public research datasets, not a clinical tool. Whether cardiologists trust a reconstructed mesh for diagnosis is a slower-moving question than whether it scores well on ACDC, and the code, posted on GitHub, is aimed at other researchers for now, not hospitals.